I take neurosymbolic AI agents from research to production — deterministic by design.

Senior Data Scientist · IDFC First Bank

vKYC AI — Research to Production

• Owned the problem framing through live deployment for RBI-compliant video-KYC, researching and productionizing a suite of biometric & anti-fraud models: speaker validation (95% recall), deepfake detection (85% recall), multi-face detection (100% recall), and PAN/PVC document verification (99% accuracy).
• Designed the fraud-signal pipeline to run inline on live vKYC calls rather than as after-the-fact review, so agents and auditors get a risk verdict within the call window instead of a batch flag hours later.
• Integrated Deepgram/Greylabs speech-to-text with transcript-based LLM validation, replacing manual call review with an automated pass that flags policy violations, script deviations, and consent-language gaps for human audit. Impact: 3,500–5,000 calls/day audited, SLA 15 min → 4 min.
• Worked cross-functionally with compliance and audit teams to keep the model suite aligned to evolving RBI video-KYC guidelines, and instrumented recall/precision dashboards per model so drift could be caught before it reached production SLAs.

Document Authenticity & Forgery Detection — Research Lead

• Led research into generative approaches (GANs, VAEs, autoencoders) for detecting tampered and synthetically forged identity/financial documents, benchmarking reconstruction-error and latent-space anomaly signals against real fraud cases sourced with the risk team.
• Designed a hybrid unsupervised/rule-based detection method — combining generative anomaly scoring with deterministic forensic checks (metadata, font, and layout consistency) — built jointly with the forensics & risk-containment teams so outputs mapped directly onto existing investigator workflows.

Document AI Extraction Platform (GenAI)

• Architected a three-tier agentic prompt-optimization system — generator, evaluator, reflector — that iteratively rewrites and scores its own prompts against held-out extraction cases, converging to 98%+ production accuracy on multi-value entity extraction across banking documents without manual prompt tuning.
• Built synthetic ground truth via multi-model consensus and LoRA fine-tuned Mistral, Qwen, GPT, and Llama variants against it, closing the labeled-data gap that would otherwise have required large-scale manual annotation.
• Established the LLM evaluation & benchmarking harness that now governs model selection bank-wide — standardizing how new open- and closed-source models are scored before they're approved for any GenAI use case. Impact: ₹30 Cr/year (~$3.6M/year).

Senior AI Scientist · StatusNeo Consulting

Legal Intelligence — Chat with Contracts (AT&T)

• Built a GenAI contract-understanding platform with chatbot Q&A over a Milvus + MongoDB store, with dynamic chunking and embedding pipelines tuned to contract-length documents rather than fixed-size splits, so retrieval stayed accurate across short amendments and hundred-page master agreements alike.
• Engineered Graph RAG on top of an extracted contract ontology — entities and clause relationships — for relationship-aware retrieval, and Agentic RAG for multi-hop query decomposition and self-correcting retrieval when the first-pass answer didn't hold up against the source clauses.
• Impact: automated contract amendment tracking that previously required manual legal review, letting the legal team query contract history and clause changes directly instead of re-reading redlines.

Claims Audit AI (Carelon)

• Led end-to-end delivery of a claims-audit platform for a major healthcare payer — owning the document-AI pipelines and annotation frameworks that turned scanned claim attachments into structured, audit-ready records.
• Built a 96% F1 attachment classifier to route incoming claim documents by type before downstream extraction, and multimodal LLM decisioning (GPT-4, Falcon, Vicuna) to adjudicate reimbursement eligibility at scale, replacing manual first-pass review by claims analysts.

Application Developer · JPMorgan Chase & Co.

• Deployed a multi-class expense categorization model (F1 88%) and liquidity/cash-flow/Nostro account forecasting (LSTM, Prophet, ARIMA) behind a FastAPI service layer, giving treasury teams same-day forecasts instead of relying on manually maintained spreadsheets.
• Migrated the Sales Dashboard onto JPMC's BRIE data lake, designing and building a custom ETL pipeline to replace the legacy data path — cut dashboard query latency by 50% and removed a recurring source of stale-data incidents.
• Led user lifecycle & subscription systems and global payment API integrations on a serverless AWS stack, owning the systems end-to-end from design through on-call support. Impact: revenue +15% (~$1,100/day).

Software Developer · Talentica Software

• Led activation/subscription systems for a user lifecycle & payments platform, owning the logic that moved a user from signup through trial to a paid subscription state.
• Built and integrated global payment API connections on a serverless AWS stack, handling multi-currency processing and retry/reconciliation logic so failed payment webhooks couldn't silently desync a user's subscription state.

Software Developer · Reflexis Systems

• Built reinforcement-learning models to auto-generate open shifts for retail workforce scheduling, replacing rule-based shift creation with a policy that adapted to store-level demand patterns.
• Built a collaborative-filtering recommendation engine (cosine similarity, Top-N) to match available shifts to the employees most likely to pick them up, reducing unfilled shifts for store managers.

Research Consultant · WorldQuant

• Formulated quantitative financial models using statistics, stochastic methods, and portfolio-management theory to generate trading alphas for the firm, working independently against WorldQuant's internal alpha-testing framework.
• Implemented statistical and basic stochastic models to search for and validate signal ideas, iterating against historical market data before submission.
• Formulated 40+ alphas that passed into production, feeding into the investment policies the firm built on top of its alpha pool.

Intern · Vitor Health

• Developed and integrated Android modules/libraries for BLE communication with connected medical devices — ECG monitors, weight scales, and blood-pressure cuffs — into the company's application.
• Revamped the app's Bluetooth settings page to make device pairing and reconnection more reliable for end users.
• Produced two explanatory videos for DHAs covering the different tests conducted by Vitor, used to onboard health-agent users to the device workflows.

Software Engineer Intern · munshiG Pvt Ltd

• Built an end-to-end dual-user app to digitalize the company's business processes — one side for the business team, one for handling clients and employees — covering data collection end to end.
• Integrated Location API, Speech API, and Google Maps API against a Firebase backend, and implemented a multithreaded system to keep data collection responsive under concurrent field use.
• Worked on table extraction from images of dot-matrix-printed bills, and contributed to the architecture design for the company's Mark 4 product, which adopted the same architecture as the Business Development app.

Control System Intern · Steel Authority of India Limited (SAIL)

• Studied the control system model for the insulation of electric cranes at SAIL's Rourkela plant, an early exposure to industrial control systems ahead of the software/ML work that followed.