A decade in the past, Picnic got down to reinvent grocery buying with a tech-first, customer-centric method. What started as a daring experiment shortly grew right into a high-scale operation, powered by steady innovation and a willingness to problem conventions.
Alongside the way in which, we’ve learnt invaluable classes about scaling expertise, fostering tradition, and driving innovation. Some have been anticipated, others have been hard-earned, and some utterly reshaped our considering.
Listed below are 10 key classes we’ve learnt—together with how the trade is tackling related challenges.
1. Constructing a tradition of innovation
Early on, we recognized that tradition and organisation are key enablers for innovation. Therefore, we emphasised autonomy by forming small groups that take end-to-end possession. We have been impressed by Amazon’s two-pizza groups, the place small, self-sufficient groups are inspired to innovate with out paperwork.
Nonetheless, we took it a step additional and constructed an “All people is an Innovator” tradition and mixed it with our present “All people is an Engineer” tradition. Enterprise operators exterior of the tech organisation can now actively contribute to the characteristic design, specification, improvement and configuration. As an illustration, we developed frameworks like Outline, Extract, Remodel, Current (DETP), the place analysts can use SQL to construct personalised suggestions with out the involvement of engineering groups. Moreover, we developed the Picnic Web page Platform that permits enterprise operators and analysts to ship app options independently from the software program improvement cycle.
The important thing metrics we optimise for are the idea-to-impact lead time and the time share of innovating and constructing vs sustaining and working. We haven’t reached our purpose but, however we’re on path in direction of all people spending not less than 80% of their time on innovation and improvement.
2. Developer expertise as a progress catalyst
Within the early days, developer expertise was a shared accountability throughout the whole tech crew. Nonetheless, shared accountability led over time to declining accountability, unclear possession, and inefficiency in decision-making. Moreover, whereas we grew as an organisation, expectations rose and complexity elevated. Thus, we rethought our method and made Developer Expertise a first-class citizen.
To deal with this, we developed a set of instruments and processes and shaped a tradition that empowers builders to work effectively, creatively, and fortunately. As an illustration, we constructed self-service instruments for all our engineers that permit them to deal with duties like setting setup, database administration, or characteristic deployment successfully.
Subsequent, we embraced a feedback-driven tradition, gathering common enter from builders to determine ache factors of their workflows and proactively handle them. One key perception from these retrospectives was that sluggish construct occasions have been a significant blocker, so we optimised our platform setting, decreasing construct occasions from 13 minutes to only 1 minute.
Past that, we monitor DevEx metrics like developer satisfaction, cycle occasions, and deployment frequency to measure the influence of DevEx initiatives and have a good time milestones to strengthen the significance of DevEx.
Comparable initiatives have been undertaken by different giant tech organisations, e.g. Google’s Bazel construct instrument that ensures high-efficiency builds and testing at scale, Netflix’s Paved Highway Philosophy, Meta’s Glean, and Stripe’s Developer-First Tradition and Working Ideas. Going ahead, we’ll monitor intently how AI-assisted coding instruments reshape the developer expertise and convey us nearer to the final word engineering expertise with high-quality code ideas, real-time concern resolutions, and automatic code refactoring.
3. The facility of platforms
Within the first few years, we centered on constructing core performance to cowl a baseline of options throughout our merchandise. Nonetheless, by 2020, we started noticing commonalities between merchandise and duplication inside them. This led us to issue out these redundancies, outline reusable patterns, and package deal them into the Web page Platform and Picnic Assist Modules.
This platform method promotes consistency, reduces code duplication, simplifies upkeep, and lowers complexity, finally bettering the general high quality of our codebase. Moreover, we built-in instruments like Error Susceptible (which we additionally open-sourced) to catch programming errors early.
Past bettering code high quality, these help modules play a vital position in main framework migrations, akin to our transition to Spring 5 (by now we’re utilizing even Spring 6), making certain clean upgrades and long-term compatibility. By offering a stable basis, these modules speed up improvement, implement greatest practices, and preserve a unified tech stack.
Moreover, our platform groups have been instrumental in standardising provisioning and deployments by Terraform, Helm and Spacelift, and accelerating our transition in direction of CI/CD by Spinnaker, Argo CD and TeamCity.
Trying past Picnic, related platform approaches have pushed effectivity at Fb (React) and Stripe (Funds), providing highly effective abstractions for complicated flows and simplified integrations at scale.
However that is only the start. Platform considering at Picnic will proceed evolving, with superior ML performance not solely powering the platform but in addition turning into a core constructing block for the following era of our expertise.
4. Scaling infrastructure for innovation
Infrastructure is a key enabler and accelerator of innovation, driving developer autonomy, automation, and scalability. By making use of Infrastructure as Code and Steady Integration throughout our stack, we automate and standardise service configuration, making certain quick, dependable, safe, and cost-efficient deployments whereas sustaining agility.
Integrating steady integration into infrastructure administration has simplified the provisioning and modification of sources, decreasing guide effort and rising effectivity. To take this a step additional, we are actually decentralising infrastructure possession throughout product groups utilizing Spacelift. This permits builders to independently handle infrastructure wants whereas making certain compliance with organisational requirements.
This mixture of IaC, automation, and developer empowerment permits us to scale infrastructure seamlessly, fostering steady innovation and operational flexibility, as demonstrated through the COVID-19 pandemic. Comparable methods have been adopted by Netflix, which has gone even additional with chaos engineering practices to stress-test resilience.
Trying forward, we purpose to additional improve our infrastructure by working in direction of predictive infrastructure administration, proactively anticipating and dynamically allocating sources to enhance efficiency and effectivity.
5. Knowledge-driven determination making
Knowledge is for us the strategic asset to innovate, resolve and function. Within the early days, information was the inspiration to help primary operations and learn to obtain operational excellence. Over time, information grew to become the driving force for strategic decision-making and innovation.
Therefore, we constructed a lakeless Knowledge Warehouse integrating information from over 300+ microservices and working 1000’s of ETL/ELT jobs and ingesting tens of millions of occasions each day for environment friendly evaluation. This enabled us not solely to analyse information, construct experiences, and get insights, but in addition to unlock our journey to construct highly effective machine studying merchandise, with out spending weeks on information cleansing. As an illustration, we constructed a set of demand forecasting instruments leveraging Temporal Fusion Transformers (TFTS), a set of customer support agent help instruments to allow service excellence, environment friendly route planning, sensible buying lists, and a lot extra.
All of that’s solely potential by leveraging our personal ML Knowledge Platform that permits not solely ML engineers however all people in Picnic to construct, practice and deploy fashions in manufacturing. Efficient studying requires suggestions and experimentation, so we constructed a scalable experimentation framework to run many A/B exams concurrently to be taught what works and what doesn’t. By combining democratised entry to information with scalable ML infrastructure and a data-first method to software program improvement, we constructed a powerful and sustainable tradition of organisational data-driven studying.
Comparable examples within the trade embody Netflix’s use of information analytics, enabling data-driven content material creation and personalised suggestions, and Uber’s evaluation of journey information, demonstrating the right way to optimise each buyer expertise and operational effectivity.
Going ahead, we’ll see many extra AI-native product options (e.g. superior steering and automation of key flows) and AI-native enterprise models (working enterprise operations that natively rely upon AI expertise).
6. Folks-centric tech
Our folks method is deeply rooted in our conviction that “Tech is a folks enterprise.” Which means that we prioritise expertise improvement, collaboration, and steady studying over inflexible constructions and conventional hierarchies. Engineers are inspired to develop, join, and contribute from day one, creating an setting the place innovation is pushed by a shared mission relatively than top-down directives.
We attempt to rent high-calibre expertise (who doesn’t? ;)) however moreover, we additionally look fastidiously into alignment with our imaginative and prescient. Therefore, we prioritise cultural match over simply expertise and have a powerful bias in direction of missionaries as a substitute of mercenaries. Studying and progress are a steady course of; for that, we now have constructed the Tech Academy that gives tailor-made coaching for each expertise degree. Our onboarding is designed as a transformative journey relatively than simply getting began on the job.
Not solely will we nurture engineering abilities, however we additionally present intensive management coaching. Now we have put intensive thought into the correct method. Our components: management allows relatively than controls. Leaders act as mentors, coaches and enablers, relatively than micromanagers. They foster bottom-up innovation, the place concepts can emerge from any degree, and domesticate a clear, trust-based tradition that ensures staff really feel valued. It’s the rationale why lots of our tech leads are engineers who’ve naturally grown into this position!
On the similar time, we prioritise well-being initiatives that help a wholesome and balanced work setting, recognising that sustained efficiency stems from each skilled progress and private well-being.
Our method combines insights from main corporations: Valve’s flat org construction, permitting staff to decide on initiatives and type groups organically; Canva’s deal with goal over perks, aligning work with staff’ values; and NVIDIA’s tradition of mental honesty, selling steady studying and transparency.
Going ahead, we’ll work in direction of a extremely adaptive organisational construction with fluid groups that repeatedly (re-)align with strategic targets, enterprise targets, and tech priorities.
7. Steady enchancment by suggestions
Suggestions is the final word superpower for studying and progress. From day one, we embedded this philosophy into our product and tech tradition, making certain that suggestions drives innovation, high quality, and steady enchancment.
On the product facet, buyer suggestions performs a central position. We actively collect insights by in-depth interviews, app evaluations, customer support retrospectives, and social media interactions, addressing each explicitly expressed and inferred wants.
Internally, we foster a innocent incident administration tradition, the place accountability is inspired with out concern. This method turns failures into alternatives for course of refinement and systematic enhancements. Nonetheless, what makes our method actually particular is the shut relationship between our tech groups and their inner prospects. Engineers, analysts, and enterprise groups work hand in hand, making certain direct and speedy suggestions, which creates an unparalleled sense of possession and alignment.
In software program engineering, we depend on structured suggestions loops, together with peer code evaluations and instruments like Error Susceptible, to take care of code high quality. Past formal processes, we encourage crew studying and information sharing by Lunch & Be taught classes and our Tech Safari programme, which supplies new joiners a 360° view of enterprise and tech. Their progress continues by our Tech Academy, providing structured studying paths tailor-made to every expertise degree.
We love velocity, however much more, we worth feedback-driven decision-making. Therefore, we foster a tradition wherein builders and enterprise work collectively to realize velocity not by reducing corners, however by delivering high-quality MVPS and utilizing future-proof structure.
Trying forward, suggestions loops will turn out to be much more autonomous and predictive. AI will monitor person interactions, detect inefficiencies, and dynamically optimise merchandise in actual time, blurring the road between iteration and innovation.
8. Balancing agility and stability
Over time, we advanced from a fast-moving startup (don’t decide, however sure, there was even a time once we ran our improvement course of with out PR evaluations ;)) to an efficient scale-up the place we discovered a brand new candy spot between agility and stability to reliably serve our greater than 1 million prospects. The early years have been characterised by a easy improvement course of, light-weight infrastructure, and excessive autonomy with ultra-short determination traces (in different phrases: thrilling, but in addition chaotic and unsustainable). Retrospectively, there’s a temptation to romanticise this time (pure innovation, limitless startup power, David vs Goliath narrative, false reminiscence of complete freedom), however in actuality, it’s a part that may’t – and shouldn’t – be maintained. As a substitute, it must be thought of a stepping stone in direction of the scaling and maturity part.
That is the part we are actually in, with formalised improvement processes, efficient governance fashions for information and infrastructure, and platforms that allow and empower fast improvement of next-generation concepts. Our principal learnings are that agility have to be structured to scale, tradition evolves (and that’s OK!), technical debt can’t be ignored, hiring shifts from generalists to specialists, autonomy wants guardrails, and scaling a crew additionally requires scaling the management. The very best corporations by no means lose their means to maneuver quick—they simply get higher at scaling responsibly.
A key realisation on this journey has been the significance of “shift left” considering—embedding high quality, safety, and reliability earlier within the improvement course of relatively than addressing them reactively. By frontloading essential checks and automating governance, we’ve considerably lowered friction down the road, permitting us to take care of velocity with out sacrificing stability. Equally, we’ve embraced the precept of “decelerate to hurry up”, recognising that investing time in structured foundations—be it well-designed structure, strong CI/CD, or clear decision-making frameworks—pays off in execution velocity and long-term agility.
Nice examples from different tech corporations embody Spotify’s mannequin of autonomous squads to foster innovation and agility, Fb‘s evolution from “Transfer quick and break issues” to “Transfer quick with steady infrastructure”, and Netflix’s migration in direction of microservices. Going ahead, we’ll transfer from merely balancing agility and stability to scaling innovation intelligently, making certain that autonomy, effectivity, and influence develop in tandem. One other fascinating path is Autonomous Innovation Networks, the place fluid, mission-driven groups dynamically type round high-impact issues, making certain agility and engagement.
9. Embracing AI-driven automation
All people in tech loves automation, however the actuality is that automation is the ultimate step in a protracted journey. Our journey started by constructing a powerful Grasp Knowledge Basis, which laid the groundwork for our first era of techniques. These techniques adopted easy rule-based determination logic, both as hard-coded Java enterprise logic or determination tree logic encoded in configuration. Whereas this method labored nicely for the primary few thousand prospects, we shortly realised it wouldn’t scale successfully.
The lacking piece was a real-time efficiency suggestions loop that would analyse system efficiency, determine enchancment alternatives, and set off modifications in actual time. To deal with this, we constructed a close to real-time Knowledge Vault alongside our Lakeless Knowledge Warehouse to offer a unified, fine-grained view of enterprise efficiency throughout all techniques. Moreover, we ensured that every one techniques had API endpoints to configure and management their logic, permitting us to develop instruments that repeatedly monitor and enhance efficiency.
For instance, we now have edge instruments that optimise inbound, fulfilment, and last-mile operations—from dynamically slotting merchandise in warehouses to demand forecasting and route planning. The primary era of those instruments was constructed utilizing Google Scripts, enabling ultra-fast improvement but in addition resulting in a shadow IT ecosystem that was tough to take care of.
To resolve this, we constructed a Platform for Edge Methods, the place this logic might be carried out in Python, the lingua franca of all analysts at Picnic. This shift permits not solely tech groups however everybody at Picnic to contribute actively to the system panorama.
The newest evolution in our automation journey is the transition from rule-based enterprise logic to AI-based determination logic. Our purpose is to maneuver each data-driven determination into ML fashions whereas holding all remaining rule-based logic in Java, Python, or TypeScript. For instance, in route planning, the cease time calculation is AI-based, whereas the shortest path calculation makes use of a heuristic algorithm in Java.
In abstract, automation isn’t nearly robots—it’s about clever, data-driven decision-making. The important thing to success is scalability, enabling a self-learning, real-time adaptive ecosystem.
10. Prioritising long-term imaginative and prescient
Now we have constructed each our enterprise and expertise with a built-to-last mindset, aligning technique with tradition and operational decision-making with sustainable long-term targets.
From the beginning, we embraced data-first considering, customer-centric reasoning, and scalable infrastructure tailor-made for agility. This method permits us to attach short-term wants with long-term imaginative and prescient, making certain that our options stay adaptable as we develop.
On the tech facet, we centered on modular and scalable structure, developer autonomy, and shut collaboration between tech and enterprise groups. This ensures that our merchandise not solely meet in the present day’s necessities however will also be prolonged and modified with ease.
One among our core ideas is to optimise not for the quickest construct to the primary launch however for the simplest path to the one centesimal iteration. We all know that true excellence comes from steady enchancment, so we optimise for long-term outcomes relatively than short-term velocity.
After all, our journey hasn’t been with out challenges. We needed to be taught the laborious approach that over-indexing on velocity results in vital technical debt, scaling with out clear enterprise priorities results in costly overgeneralisation, and fast hiring whereas ignoring tradition results in misalignment and requires painful correction.
Trying forward, we’re exploring promising approaches akin to AI-driven operational effectivity (e.g., predictive scaling and AI for engineering), hyper-personalised expertise retention and tradition scaling, and additional decentralisation, the place engineers collaborate with AI brokers because the norm.
That mentioned, we’re absolutely conscious that AI shouldn’t be a silver bullet. It comes with limitations, dangers, and moral challenges. Not each breakthrough must be AI-driven—there are lots of different paths to innovation, and we consider AI must be used the place it provides actually distinct worth.
The teachings learnt over the previous decade are usually not simply milestones—they’re the inspiration for what comes subsequent. And one factor is evident: the journey was, is, and can by no means be linear. From embracing uncertainty to scaling with agility, the previous ten years have strengthened our sturdy perception that buyer obsession, relentless innovation, and adaptive management are key to long-term success. As we glance forward, the following wave of grocery tech innovation will demand even larger velocity, resilience, and collaboration, reminding us that the best breakthroughs are nonetheless to come back.
