Production AI —
not just notebooks.

The Problem

Financial analysts spend days compiling research from 10+ scattered sources — market data, news, SEC filings, fundamentals. Manual work doesn't scale and introduces inconsistency.

The Solution

Designed and deployed a deterministic investment decision engine — a 7-step async LangGraph pipeline that generates auditable BUY/HOLD/SELL recommendations in under 30 seconds. GPT-4o is constrained to explanation-only: zero influence over the decision, ensuring fully reproducible and auditable outputs with zero hallucination risk on recommendations.

Engineering Decisions

Normalized multi-factor scoring engine (Technical 0–25, Fundamental 0–40, Sentiment 0–15) with dynamic max adjustment — missing data excluded from denominator rather than penalised, eliminating false SELL signals. Conflict-aware decision logic: signal disagreement (variance > 0.15) overrides to HOLD in 35–55% ambiguity band, reducing false conviction. 5-factor confidence model (data completeness, signal agreement, volatility, news uncertainty) blocks recommendations below 20% confidence. Redis TTL 900s caching reduces repeated query latency from ~15s to <100ms.

Business Value

Any firm paying analysts $80k+/year offsets that cost with sub-30s deterministic decisions. Portfolio ranking mode supports 2–8 tickers with proportional capital allocation. The same architecture extends to portfolio monitoring, earnings alerts, and competitive intelligence at near-zero marginal cost per use case.

Key Results

• BUY/HOLD/SELL decision generated in under 30 seconds
• Repeated query latency: ~15s → <100ms via Redis caching (10× API cost reduction)
• Zero LLM hallucination risk — GPT-4o explains decisions, never makes them
• Portfolio ranking mode: 2–8 tickers with proportional capital allocation
• Live on AWS EC2 with Nginx domain routing and Docker Compose

Stack

The Problem

Enterprise knowledge locked in PDFs, Word docs, and internal wikis means employees waste hours on search instead of work. Generic chatbots hallucinate without grounding.

The Solution

Engineered a production RAG system: PDF ingestion pipeline (1000-char chunks, 200-char overlap, OpenAI text-embedding-3-small), ChromaDB vector store, MMR retrieval, and GPT-4 generation with real-time streaming. Hallucination guardrails block LLM calls when retrieval confidence is too low. Deployed on AWS EC2 with HTTPS via Let's Encrypt.

Engineering Decisions

MMR (Maximal Marginal Relevance) retrieval over plain cosine similarity — fetches a larger candidate pool and re-ranks for relevance and diversity, preventing near-duplicate chunks reaching the LLM. Rule-based query classifier (no LLM cost) detects factual/complex/ambiguous/keyword intent and adjusts top_k and fetch_k accordingly. Three API cost guards: rate limiting (20 req/min via slowapi), 500-char input cap, tiktoken token budget — all fire before any LLM call. Every answer uses [SOURCE N: filename — Page: X] citation format — LLM is forced to cite, never invent.

Business Value

Dual deployment: Streamlit chat UI for business users, FastAPI Swagger UI at /api/docs for engineering integration — one system serves both audiences. The citation format makes every answer auditable — critical for regulated industries. Cost guards and query classifier make this viable at production scale without API spend ballooning.

Key Results

• 87% retrieval accuracy on benchmark query set
• Zero hallucination on document-grounded queries — guardrails return fallback instead of guessing
• MMR retrieval prevents near-duplicate context from polluting LLM answers
• Two interfaces: Streamlit (users) + FastAPI Swagger (developers)
• Live on AWS EC2 with Nginx reverse proxy and HTTPS (Let's Encrypt)

Stack

The Problem

Single-model forecasts fail to capture both long-range trends and short-range seasonality simultaneously, causing inventory errors that cost retailers millions annually.

The Solution

Architected an LSTM + Prophet ensemble — LSTM captures complex temporal dependencies, Prophet handles seasonality and holidays. Trained on 2 years of multi-store data, deployed as a SageMaker endpoint with automated weekly retraining triggered from S3. Model stays current without manual intervention.

Key Results

• 4.2% RMSE across 45 product categories — 18% improvement over single-model baselines
• Trained on 500,000+ retail transactions across 2 years of multi-store data
• 8-week forecast horizon with automated weekly SageMaker retraining from S3
• Ensemble captures both temporal complexity (LSTM) and seasonality/holidays (Prophet)

Stack

The Problem

Acquiring new telecom customers costs 5× more than retaining existing ones. Without early warning signals, at-risk customers leave before retention campaigns can reach them.

The Solution

Engineered 21 customer behaviour features plus AWS Comprehend NLP sentiment signals extracted from 7k+ customer support tickets — expanding the feature space with real customer voice data. End-to-end MLOps: S3 ingestion → Lambda ETL → SageMaker training → real-time inference endpoint. Automated retraining via CloudWatch + EventBridge on data drift detection. SHAP waterfall plots make every prediction auditable for non-technical retention teams. MLflow tracks all experiments with full reproducibility.

Key Results

• 91% classification accuracy, 89% F1 score
• AWS Comprehend NLP sentiment from 7k+ customer support tickets
• Full MLOps pipeline: S3 → Lambda ETL → SageMaker → real-time inference endpoint
• CloudWatch + EventBridge automated retraining on data drift
• SHAP waterfall plots for every individual prediction

Stack

The Problem

Insurance companies process 50,000+ claims annually. Manual review is slow, expensive, and inconsistent — letting fraudulent claims slip through while delaying legitimate ones.

The Solution

Designed a dual-model CV pipeline: YOLOv8 localises vehicle damage in real time, ResNet-50 classifies severity and flags fraud indicators at 96.7% accuracy. Real-time claim ingestion via AWS Kinesis, serverless processing with Lambda, and fraud scores stored in RDS — enabling SQL-based analyst queries on top of the pipeline. FastAPI serves the inference endpoint at 50ms average latency.

Key Results

• 96.7% fraud detection accuracy on 50,000+ claim images
• 50ms average inference latency via FastAPI REST API
• Real-time ingestion via AWS Kinesis + Lambda serverless processing
• Fraud scores stored in RDS — analysts query results directly via SQL
• Dual-model pipeline: YOLOv8 (detection) + ResNet-50 (classification)

Stack

The Problem

Film studios spend millions producing trailers with no way to quantify emotional impact before release. Traditional test screenings are expensive, slow, and limited to small samples — there's no second-by-second signal showing where audiences peak in excitement, disengage, or feel fear.

The Solution

CineNeuro uses Meta FAIR's TRIBE v2 brain encoding model to predict fMRI activations across 20,484 cortical vertices for every second of a trailer. Three feature extractors process the video simultaneously — V-JEPA2 for visual content, Wav2Vec-BERT for audio waveforms, and Llama 3.2 for transcript embeddings. These feed TRIBE v2, and the predicted activations are mapped to 5 emotion channels (excitement, fear, joy, suspense, boredom) across 7 functional brain regions using neuroscience-based weights.

Engineering Decisions

Two-tier deployment strategy: a g4dn.xlarge GPU instance (Tesla T4, 15GB VRAM) handles inference on-demand at $0.53/hr, then terminates. A t3.micro serves the React dashboard 24/7 at near-zero cost using pre-computed JSON results (<100ms load time). The three extractors total 12.89GB (V-JEPA2 4.14GB + Wav2Vec-BERT 2.32GB + Llama 3.2 6.43GB) — barely fitting in 15GB VRAM, requiring careful sequential model loading. Multi-stage Docker build (Node 18 + Python 3.11-slim) keeps the production image at ~500MB with no ML weights. Submitted PR #20 to Meta FAIR's TRIBE v2 — added int8 compute_type fallback for WhisperX on non-CUDA devices. Before the fix: hardcoded float16 caused ValueError on Apple Silicon and CPU-only Linux. After: the full pipeline runs everywhere. CUDA path unchanged, <1% WER degradation on transcription accuracy. CLA signed, checks passing.

Business Value

Second-by-second emotion maps identify the 3 strongest and 3 weakest trailer moments with human-readable explanations — directly actionable edit intelligence. Audience persona simulation (Action Lovers, Romance Fans, Horror Enthusiasts) shows differentiated engagement curves. Competitive benchmarking against 5 iconic trailers (Oppenheimer, Avengers Endgame, Inception, Interstellar, The Dark Knight) gives studios a quantified baseline. Full analysis costs ~$0.50 in GPU time vs thousands for a traditional test screening.

Key Results

• 20,484 cortical vertices predicted per second of video across 7 functional brain regions
• 5 simultaneous emotion channels: excitement, fear, joy, suspense, boredom
• 53–224 predicted segments per trailer (74–93% kept after quality filter)
• Pre-computed demo results served in <100ms from disk
• Full 2-minute trailer analysis in ~55 minutes on Tesla T4 GPU
• Monthly hosting cost: ~$0 (t3.micro free tier + on-demand GPU only when needed)
• PR #20 submitted to Meta FAIR TRIBE v2 — int8 compute_type fallback for WhisperX on non-CUDA devices. CLA signed, checks passing.

Stack