Aditya Mahakali

Highlights

• Unified dense + structured retrieval for operational queries.
• Designed for reliability under enterprise constraints.

Architecture

• Ingestion: document normalization, chunking, and metadata enrichment.
• Indexing: hybrid dense/sparse signals + structured fields for filtering.
• Retrieval: query parsing -> hybrid retrieval -> ranking -> context assembly.
• Generation: answer synthesis with citations and guardrails.

Model + quality work

• Retrieval tuning: hybrid weighting, filters, and ranking features based on evaluation outcomes.
• Prompting: citation-first answering and refusal patterns for low-confidence contexts.
• Embeddings: model selection + chunk sizing tuned to networking-style documents.

Evaluation

• Offline: retrieval recall/precision on curated query sets.
• Online: user feedback loops and failure-mode review (hallucination, mismatch).

Deployment

• API-first backend with observability hooks (latency, retrieval hit-rate).
• Security: RBAC-ready integration and input sanitization patterns.

Full stack

Highlights

• Intent-based flows with grounded answers for self-service.
• Deep linking from answers into product/knowledge pages.

Architecture

• Channel/UI -> assistant orchestration -> retrieval service -> LLM answer synthesis.
• Chunking: parent-child strategy for better grounding on long docs.
• Retrieval: hybrid semantic + lexical with query rewriting and filters.

Model + quality work

• Conversation quality: prompt patterns for tool use, grounded answers, and consistent tone.
• Retrieval quality: parent-child chunking + query rewriting heuristics for intent-specific recall.
• Safety: prompt-injection aware prompting + constrained tool behavior.

Evaluation

• Conversation-level QA: groundedness checks and citation coverage.
• Task success tracking via intent completion and fallback rates.

Deployment

• Containerized services for backend + UI; environment-specific configs.
• Safety: prompt-injection aware prompting + restricted tools/actions.

Full stack

Highlights

• Metadata summarization to improve searchability and ranking signals.
• Assistant layer on top of search for guided drug/product queries.

Architecture

• Ingestion: parse FDA documents -> sectioning -> metadata extraction.
• Summarization: generate structured product summaries as searchable metadata.
• Search: rank using lexical + metadata fields; assistant uses search as tool.

Model + quality work

• Summarization: structured prompting + schema validation to keep metadata consistent.
• Search tuning: field boosts and ranking rules leveraging summary metadata.
• Risk control: safer phrasing + guardrails for regulated/medical content.

Evaluation

• Search relevance evaluation on representative queries (A/B on metadata).
• Assistant evaluation: answer grounding and incorrect-drug risk checks.

Deployment

• Search service + assistant APIs; guarded outputs for regulated content.
• Monitoring: drift checks on summarization schema adherence.

Full stack

Highlights

• Business users query data safely without writing SQL.
• Guardrails for RBAC + prompt-injection resistance.

Architecture

• Metadata dictionary: automated schema understanding + synonyms.
• Pipeline: NL -> intent/slots -> SQL draft -> validation -> execution -> explanation.
• Safety: RBAC enforcement + allowlisted tables/columns + query constraints.

Model + quality work

• Schema linking: metadata dictionary + synonyms to reduce ambiguity and improve grounding.
• Accuracy tuning: curated few-shot examples + error-driven prompt iteration.
• Safety: allowlists, query validation, and injection defenses before execution.

Evaluation

• SQL accuracy: execution match + result correctness on benchmark queries.
• Security tests: injection attempts, privilege escalation, schema leakage.

Deployment

• Microservice API with audit logs for generated SQL and user identity.
• Latency tuning: caching metadata and reusing compiled prompts/templates.

Full stack

Highlights

• Improves retrieval coverage for sparse or ambiguous queries.
• Designed to be testable with offline evaluation sets.

Architecture

• Query -> hypothetical doc generation -> embedding -> retrieval -> rerank -> answer.
• Strict separation between retrieval augmentation and final answer generation.

Model + quality work

• HyDE prompting: controlled hypothetical generation to boost retrieval recall without hallucinated final answers.
• Embedding + rerank: model selection and reranking strategy tuned for medical relevance.
• Safety: refusal patterns and disclaimers to avoid medical advice overreach.

Evaluation

• Compare baseline RAG vs HyDE on retrieval metrics and answer quality.
• Failure analysis: misleading hypotheticals and unsafe medical outputs.

Deployment

• Batch evaluation harness + lightweight API wrapper for demos.
• Safety: disclaimers and refusal behavior for medical advice.

Full stack

Highlights

• End-to-end automation: fetch -> filter -> summarize -> publish.
• Separation of agents: retrieval, summarization, quality checks.

Architecture

• Source ingestion -> relevance filter -> multi-agent summarization -> final editor pass.
• Tooling: per-agent prompts and deterministic output schemas.

Model + quality work

• Quality: agent role prompts + schema enforcement to reduce drift and improve consistency.
• Cost/latency: step budgeting and retry strategy for flaky sources/tools.
• Faithfulness: citation-first summaries and checks for unsupported claims.

Evaluation

• Quality rubric: coverage, faithfulness, and novelty of takeaways.
• Latency/cost tracking per agent step to optimize throughput.

Deployment

• Scheduled runs + artifact storage for summaries and references.
• Retry strategy for flaky sources and tool failures.

Full stack