Research to production
Ambiguous problems taken from framing through deployment and org-wide adoption, not just notebooks.
Neurosymbolic AI • Research to Production • Deterministic Agents
I take neurosymbolic AI agents from research to production — deterministic by design.
Full-stack AI scientist with 7+ years and a Dual Degree in M.Sc. (Hons.) Mathematics + B.E. (Hons.) Electrical & Electronics — I fuse neural models with symbolic structure (rule-based logic, knowledge graphs, ontologies, evaluation harnesses) to make agentic AI systems deterministic, auditable, and production-safe. Across banking, healthcare, and enterprise I own problems end-to-end — framing, research, architecture, deployment, and adoption.
- 7+ years across banking, healthcare & enterprise AI
- Rule-based logic, knowledge graphs & ontologies wrapped around GenAI, CV, NLP, speech, and RL
- ₹30 Cr/year (~$3.6M) measurable impact
Bengaluru, India
SS
7+
years experience
95–99%
production metrics
Current focus
Banking, financial services & enterprises
7+
years experience
₹30 Cr/yr
measurable impact (~$3.6M)
95–99%
production accuracy, shipped models
5
companies, end-to-end ownership
About
Neurosymbolic AI, built for determinism.
I'm Shashank, a full-stack AI scientist with a Dual Degree in M.Sc. (Hons.) Mathematics + B.E. (Hons.) Electrical & Electronics from BITS Pilani. My throughline is neurosymbolic AI — pairing neural models with symbolic structure (rule-based logic, knowledge graphs, ontologies, evaluation harnesses, agentic control loops) so agentic systems behave predictably in production, not just in the lab. Over 7+ years across banking, healthcare, and enterprise, I've shipped nearly every class of AI problem — GenAI (LLM fine-tuning with LoRA, agentic prompt optimization, Graph & Agentic RAG, multimodal decisioning, synthetic data via multi-model consensus), computer vision, NLP, speech, document AI, time series forecasting, recommendation, and reinforcement learning.
I own problems end-to-end — framing, research, architecture, deployment, and org-wide adoption — and I measure success the same way the business does: in production metrics and delivered impact.
Neurosymbolic by design
Neural models paired with symbolic structure — rule-based logic, knowledge graphs, ontologies — so agentic systems behave the same way twice, not just most of the time.
Full-stack ownership
Android, web, backend, and ML in one hand — I ship the whole surface, not just the model.
Measurable impact
Every project tied to a concrete production metric — recall, accuracy, latency, revenue.
Experience
Seven years, five companies, one thread: ship it.
Senior Data Scientist · IDFC First Bank
vKYC AI — Research to Production
- Researched and productionized biometric & anti-fraud models for RBI-compliant video-KYC — speaker validation (95% recall), deepfake detection (85% recall), multi-face detection (100% recall), PAN/PVC verification (99% accuracy).
- Integrated Deepgram/Greylabs speech-to-text with transcript-based LLM validation, replacing manual call review. Impact: 3,500–5,000 calls/day, SLA 15 min → 4 min.
Document Authenticity & Forgery Detection — Research Lead
- Researched generative approaches (GANs, VAE, AE) for forgery detection, designing a hybrid unsupervised/rule-based method with forensics & risk-containment teams.
Document AI Extraction Platform (GenAI)
- Architected a three-tier agentic prompt-optimization system (generator, evaluator, reflector), converging to 98%+ production accuracy without manual prompt tuning.
- Built synthetic ground truth via multi-model consensus, LoRA fine-tuned Mistral, Qwen, GPT and Llama, and established the LLM evaluation & benchmarking harness governing model selection bank-wide. 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 Milvus + MongoDB; engineered Graph RAG and Agentic RAG for multi-hop, self-correcting retrieval. Impact: automated contract amendment tracking that previously required manual legal review.
Claims Audit AI (Carelon)
- Led end-to-end delivery of a claims-audit platform — document-AI pipelines, annotation frameworks, a 96% F1 attachment classifier, and multimodal LLM decisioning (GPT-4, Falcon, Vicuna) for reimbursement adjudication at scale for a major healthcare payer.
Application Developer · JPMorgan Chase & Co.
- Deployed multi-class expense categorization (F1 88%) and liquidity/cash-flow/Nostro forecasting (LSTM, Prophet, ARIMA) via FastAPI; migrated the Sales Dashboard to JPMC's BRIE data lake with a custom ETL pipeline — 50% lower dashboard latency.
- Led user lifecycle & subscription systems and global payment API integrations on a serverless AWS stack. Impact: revenue +15% (~$1,100/day).
Software Developer · Talentica Software
- Led activation/subscription systems and global payment API integrations on a serverless AWS stack for a user lifecycle & payments platform.
Software Developer · Reflexis Systems
- Built RL models for open-shift generation and a collaborative-filtering recommendation engine (cosine similarity, Top-N) for retail workforce scheduling.
Selected work
Case studies from research to production.
Agentic Document Extraction Platform
A three-tier agentic prompt-optimization system — generator, evaluator, reflector — converging to 98%+ accuracy on multi-value entity extraction across banking documents, without manual prompt tuning.
- LoRA fine-tuning: Mistral, Qwen, GPT, Llama
- Multi-model synthetic consensus data
- ₹30 Cr/year (~$3.6M/year) impact
vKYC Biometric & Anti-Fraud Suite
RBI-compliant video-KYC models taken from problem framing to live deployment: speaker validation, deepfake detection, multi-face detection, and document verification.
- 95% speaker validation recall
- 100% multi-face detection recall
- SLA cut 15 min → 4 min
Legal Intelligence — Chat with Contracts
A GenAI contract-understanding platform with chatbot Q&A and semantic search — entities and relationships extracted into a contract ontology for relationship-aware retrieval (Graph RAG) and Agentic RAG for query decomposition and self-correction.
- Milvus + MongoDB semantic search
- Dynamic chunking & embedding pipelines
- Automated contract amendment tracking
Claims Audit AI
An end-to-end claims-audit platform for a major healthcare payer — document-AI pipelines, annotation frameworks, and multimodal LLM decisioning for reimbursement adjudication at scale.
- 96% F1 attachment classifier
- Multimodal decisioning: GPT-4, Falcon, Vicuna
- Production QA gates on entity extraction
Nityarth — CRM & Payroll Platform
A production CRM and payroll web application built end-to-end for a client — SPA frontend, API backend, and a full CI/CD pipeline shipping every change straight to production.
- Custom domain with automated HTTPS
- Jenkins-driven CI/CD deployment pipeline
- Live at nityarth.in
Skills
The stack behind the impact.
GenAI & LLMs
GPT, Claude, Llama, Mistral · LoRA/QLoRA/RLHF fine-tuning & prompting
Agentic & Neurosymbolic Systems
Multi-agent systems, MCP · RAG, Graph RAG, Agentic RAG · ontology & knowledge-graph design · rule-based hybrid reasoning · LangChain/LangGraph/LlamaIndex
ML & deep learning
PyTorch · NLP, computer vision, multimodal transformers · document AI
Infra & MLOps
Docker, Kubernetes, CI/CD · AWS Bedrock/SageMaker, Azure, Vertex AI · Spark, Hadoop, SQL/NoSQL
Languages & systems
Python, Java · system design · vector DBs & semantic search (Milvus, FAISS)
Beyond the lab
Range, off the clock.
Outside of shipping models, I write, make films, sketch, compete, and travel for months at a time. Different disciplines, same instinct: frame the problem, commit, finish the thing.
Competitor
Cricket, football, badminton, MMA.
Voice
Public speaking and oratory.
Musician
Guitar, mouth organ, and a bit of singing.
Traveller
Long-form travel — months on end, off the beaten schedule.
Writing
Notes, essays, and half-formed arguments.
Medium · Essay
Marketers should try to influence narcissists
Medium · Essay