Prabakaran Chandran

Owned product work on an agentic analytics and enterprise data platform end to end, from algorithm research to system building and scaling, and evaluated vendor and technology stacks for the product and infrastructure roadmap.

• Agent Harness: Built an agent harness on LangGraph and DSPy, extensible across use cases, with durable execution for long-running agent tasks, semantic, episodic and procedural memory, tracing and observability, and GEPA (reflective prompt evolution) self-optimization of DSPy agent programs without model training, served on Kubernetes and scaled as a product.
• Context Engineering: Constructed the ontology, semantic and context layers algorithmically from data assets and documents using long-context deep-research agent swarms; prototyped a neurosymbolic layer over telemetry, runbooks and incident history.
• Data and Analytics Agents: Built grounded NL-to-SQL (text-to-SQL) on that semantic layer without model training, and sandboxed agents that build data applications and ad hoc analytics reports over the same ontology and context layer.
• Evaluation and Verification: Evaluated agents with personality-based and other simulations, statistical tests, grounded verification and LLM-as-a-judge, and prototyped a coding agent that builds and tests sub-agents in a sandboxed environment with task graders.
• Research Prototypes: Prototyped IRM-style (invariant risk minimization) continual learning that kept deployed agents stable through environment drift, tabular foundation models for network security and behavior, and hypergraph reasoning with swarm coordination.

Led the founding AI team in document intelligence and long-context evidence synthesis for clinical, legal and insurance work, owning architecture and implementation through beta and training front-end developers and new graduates on LLM and vision-language solutions.

• LLM Post-Training: Curated around 100K annotated medical documents and synthetic data, adapted Qwen-2.5 with continued pre-training, LoRA, SFT and preference alignment (DPO, ORPO, GRPO), trained small reasoning models for legal and insurance work, served on vLLM for testing and adoption.
• Evidence-Grounded Generation: Ran long-context, long-horizon evidence mining and synthesis over thousands of electronic and scanned health records, producing chronologies and evidence reports of about 50 pages each, served from backend microservices.
• Document AI: Engineered a routed multimodal extraction workflow for scanned PDFs and EHR exports: document-type routing, vision-language model extraction, schema reconciliation and ICD-10 mapping, after benchmarking OCR engines.
• Medical Question Answering: Built RAG with vector search over medical embeddings and late-interaction retrieval (ColBERT, ColPali), and clinical reasoning that inferred malpractice, side effects and misrepresentation from extracted records.

Led applied research in AIOps and data observability (anomaly detection, root-cause analysis and predictive analytics for IT operations) with cross-functional engineering teams, and deployed an on-premise log analysis and insights dashboard. (FastAPI)

• LLM Post-Training: Owned post-training of models up to 32B on multi-GPU nodes: continued pre-training of Phi-3 and Mistral, DPO and ORPO for function calling, and GRPO for grounded reasoning over multi-system logs, traces and metrics, judged by LLMs.
• Anomaly Detection: Chose few-shot contrastive learning over LLM prompting for anomaly detection in failed-job logs (F1 0.895 on critical error signals), and trained log embeddings for root-cause clustering, prototyping graph attention networks. (SetFit)
• Agents in Production: Shipped forCE, a multi-agent system that parses cloud data-integration failure logs, explains which thread failed and why, and retrieves fixes, making around 60% of P2 and P3 support solutions self-served. (LangGraph, RAG, FastAPI)
• Knowledge-Work Agents: Implemented agents for customer support, ERP and analytics, including a multi-agent system for NL-to-SQL, presentations and documentation, after knowledge graph extraction experiments on CRM conversations.

Built the Earth observation and geospatial ML practice for seafood supply-chain traceability from scratch, owned the traceability engine end to end, and contributed product scope, roadmap and budget plans to the entrepreneur-in-residence mission after coastal field visits.

• Computer Vision at Scale: Mapped over 1,200 hectares of aquaculture ponds from Sentinel-2 at F1 0.92 with YOLOv8-guided, fine-tuned Segment Anything and a newly derived loss, after benchmarking DeepLabv3, SegFormer and MaskFormer.
• Performance Optimization: Deployed the segmentation models on AWS SageMaker with serverless, asynchronous and batch inference, cutting the MLOps pipeline’s latency 37%, and automated Sentinel-1/2 zonal-statistics data pipelines. (Dask, Xarray)
• Forecasting: Modeled fish migration and forecast ocean conditions for potential fishing zones with a species-conditioned spatio-temporal deep learning model, and assessed overfishing with neural latent Dirichlet allocation in a proof of concept. (Ray)

Rose from intern to team lead in three and a half years: set technical direction for delivery teams across business units, built winning RFP responses, and led hiring and training of new cohorts and client teams, presenting to Fortune 100 leadership.

• Supply Chain Optimization: Architected a USD 8M company-wide sales and operations planning modernization for a Saudi petrochemical firm, from the winning proposal and team training to demand-sensing and supply-network optimization models.

Delivered decision science (forecasting, causal measurement, optimization, simulation) for Fortune 500 clients in energy, manufacturing, consumer goods and retail, from applied research to product building and implementation, working directly with client teams.

• Forecasting: Beat an energy client’s incumbent solar forecast for power-trade bids: PySpark data pipelines statistically corrected weather forecasts before Bayesian-tuned ML models with conformal intervals, shipped with a what-if dashboard. (Ray, MAPIE)
• Experimentation: Measured promotion impact with propensity-score matching and sales drivers with mixed-effects and regularized market-share models, A/B tested a construction budget model against traditional heuristics, and quantified advertising effects with an agent-based simulation driven by an expert-elicited Bayesian network. (Python, SQL, R)
• Computer Vision: Cut manual QA of 3D-printed reactor and jet-engine parts from days to hours with instance segmentation and object detection in a three-tier web app, and combined physics-informed neural networks with Pix2Pix conditional GANs to search CAD images and process parameters for root causes of overheating defects. (React)
• Geospatial Analytics: Engineered geospatial pipelines over GPRS and beacon logs to map crowd mobility around points of interest, related those points to population demographics with tessellation analysis, and ran Voronoi site selection for a smart-city developer.
• Pricing Optimization: Engineered NLP-powered price prediction for energy, oil and gas clients from tabular data, text extraction and a knowledge graph, built price-elasticity models for dynamic pricing, and optimized anti-dumping export prices in Gurobi.

• Time-Series Forecasting: Developed a pharmaceutical sales forecasting model using a state-space model with Kalman filtering.

• Causal Inference: Estimated campaign treatment effects with a hierarchical causal state-space model over 276,484 observational baskets from 2,500 households, finding +8.2% lift for one campaign that a naive estimator overstated 2.4x (E-value 1.37).
• Causal Representation Learning: Separated platform-induced engagement from user preference in KuaiSAR logs with HMV-CRL, a two-encoder causal representation learner (MCC 0.944 and 0.713; disentanglement index 82.5 under simulated policy shift).
• Continual Learning: Derived exemplar-free methods from one causal assumption: NSM-LW reached 76.58% on Split CIFAR-10, 31.3 points above DER++ with a 200-exemplar buffer, and a variant beat published replay by 28.4 points on Split CIFAR-100.
• World Models: Built CF-JEPA, a counterfactual JEPA (joint-embedding predictive architecture) world model designed to infer hidden block masses from video by event-weighted pooling and roll out counterfactual latents with a Koopman-linear predictor. (PyTorch)
• Self-Supervised Audio: Designed Factorize-and-Compose Audio-JEPA (BirdCLEF 2026), a self-supervised model predicting a field recording’s embedding as the sum of its species’ clean-clip embeddings, over 234 species and about 128K unlabeled windows.
• Multi-Agent Reinforcement Learning: Researched zero-shot coordination in cooperative multi-agent RL, porting the Overcooked-V2 environment to a standalone JAX package (23 layouts, 7 passing tests) without the JaxMARL dependency.
• Evaluation and Verification: Predicted coding-agent failure from 4,495 SWE-agent trajectories with instance-grouped 5-fold cross-validation, reaching 0.639 within-instance ranking accuracy (length alone 0.599) after removing a trace-length leak.