Archit Roy

About

I build production data-science systems at the point where models meet products: dynamic pricing, behavioural prediction, generative AI, evaluation, APIs and the operational machinery that keeps them trustworthy.

My default language is Python. I enjoy going from first-principles research and exploratory notebooks to data contracts, failure modes, observability and maintained services. I care as much about proving when a model should not act as I do about improving its score.

Outside work, I am drawn to ambitious stories, space and long timelines. I like learning subjects deeply enough to rebuild their history, understand why the current approach won and see what still remains unsolved.

Experience

Selected professional systems

• DS-NEMO: A review-summarisation system developed through classical NLP, fine-tuned transformers and LLMs, then productionised around evaluation and low-latency serving. Associated with a 46% reduction in review clicks and a 12% reduction in checkout-session duration. The API handled 200K+ daily requests with an uncached p99 of 17 ms. Read technical case study.
• AIB-TANISHA: A policy-aware multilingual voice campaign for giving the platform's quiet majority of friction-free trips a fair chance to be represented in app-store reviews. Produced a concise, versioned conversation specification and adversarial scenario matrix for the campaign. Read technical case study.
• DS-OSIRIS: A production propensity system and controlled experiments that separated predictive ranking from intervention impact. The top decile converted at about 3.16× baseline; experiments showed that a strong ranker does not guarantee a useful intervention. Read technical case study.
• DS-PIXEL: An end-to-end generative image pipeline with object-consistency checks, aesthetic scoring and marketplace experiments. Produced directional CTR and ranking improvements across city experiments while filtering many harmful edits. Read technical case study.
• DS-CIPHER: An LLM-assisted support system improved through manual false-positive analysis, stakeholder adjudication and duplicate-safe processing. Reached roughly 51% automated resolution across eligible flows at about 94% audited accuracy. Read technical case study.
• DS-WATCHDOG: A daily evaluator that answers a simple question: did a business metric genuinely move, or are we looking at normal variation or incomplete data? Converted an exploratory workbook into a replay-tested daily service with explicit data-quality states. Read technical case study.
• DS-PIGEON: A common reporting layer that replaced scattered mailers with gated, previewable and consistent operational communication. Created a reusable last mile for analytical products without duplicating their business logic. Read technical case study.
• PRICING: Base fares, behavioural ranges, dynamic multipliers, product economics, interventions and change management treated as one ecosystem. Made high-impact pricing changes faster, more explicit and easier to trace without a disruptive rewrite. Read technical case study.
• DS-ALERTS: A central reliability layer for execution health, infrastructure failures and database integrity across production data-science and data-engineering services. Created a common operational safety net across the DS–DE production estate. Read technical case study.

Technical release notes

• Serving-path optimisation: A focused reliability and performance pass on Nemo’s high-volume, read-heavy API. Read release case study.
• Goa base-fare automation: A maintained Goa pricing path that replaced file hand-offs with automated calculation, layered fallbacks and regular publication. Read release case study.
• Variant-aware base fare: A base-fare release that prefers car-variant acquisition data while retaining measurable fallback coverage. Read release case study.
• Limited and unlimited kilometre pricing: A recommendation-contract release that made kilometre-plan flexibility part of pricing identity. Read release case study.
• Asset-level pricing control: A bounded, auditable car-level multiplier with neutral onboarding and controlled bulk changes. Read release case study.
• Reality, Illusion and one portfolio: A personal website with a plain professional record, an interactive retro desktop and one shared content model behind both. Read release case study.

Technical skills

LANGUAGES

Python, SQL, C++

ML + STATS

XGBoost, Scikit-learn, SHAP, Pandas, NumPy, Experimentation

NLP + GENERATIVE AI

Gemini, Transformers, T5, BART, PEGASUS, RAG, LLM evaluation

SYSTEMS

Flask, Gunicorn, SQLAlchemy, Docker, Kubernetes, Kafka

DATA + CLOUD

BigQuery, MySQL, GCP, Vertex AI, ETL pipelines

OBSERVABILITY

Langfuse, OpenTelemetry, New Relic, Last9, Jenkins, ArgoCD

Education

Vellore Institute of Technology — B.Tech, Information Technology · CGPA 9.09 / 10.0 (2021 — 2025)

Kunskapsskolan — CBSE · Grade 12: 94% · Grade 10: 91% (SCHOOL)

Certifications

• AWS Certified Cloud Practitioner (CLF-C02)
• Microsoft Power Platform App Maker Associate (PL-100)

Outside Work

• STORIES: I love stories with large worlds and real interior lives—science fiction, fantasy, anime, films, series and books. I am interested in how a story builds rules, earns emotion and changes the scale at which you see ordinary life.
• SPACE: Space is my antidote to the smallness of a bad day. Thinking across planets, black holes and billions of years restores proportion without making the present feel meaningless.
• GROWTH: To live is to grow. I like rebuilding ideas from their history, writing down what failed and returning to a problem with a better model of it.
• SYSTEMS: I am drawn to systems with clear ownership, recovery, observability and feedback—technical or otherwise. Reliability is what turns an interesting idea into something people can trust.

This page intentionally contains no interface, animation or visual trick. It is the record. Shout out to the real legends whose shoulders we stand on today.