Project Showcase

TTOs sit on backlogs of papers and need first-pass commercialization verdicts. Manual review takes experts days per paper; single-prompt LLM evaluation invents market figures and readiness scores nothing can be traced to.

Three runtime planes: fast Next.js 14 API routes, a durable Inngest pipeline running five agents in isolated step.run() blocks with 3× retries, and a Pusher + polling notification layer. Neon Postgres via Prisma persists one schema-validated model per agent output.

Hallucinated market data; multi-minute LLM jobs on serverless timeout limits; partial failures mid-pipeline (search down, model 503s); keeping five agents' outputs mutually consistent; unreliable LLM output shapes.

Closed-world retrieval (market synthesis can only cite URLs actually retrieved via Tavily), a ref-menu restricting the pitch builder to enumerated upstream facts, adversarial critique agents at three points with bounded one-shot regeneration, skip-don't-fail semantics for non-core stages, and dual Zod + Gemini responseSchema enforcement.

Small teams without dedicated creatives can't sustain consistent multi-channel marketing. Generic LLM output drifts off-brand and ignores the product data that actually differentiates the offer.

Event-driven backend: CSV/Excel ingestion → rule-driven attribute extraction → fine-tuned Llama-based generation served over FastAPI on GPU-optimized inference endpoints, with transformer-based tone modulation per channel.

Grounding generated copy in ingested product attributes, controlling tone per channel from one model, and keeping inference latency interactive on GPU budgets.

Deterministic preprocessing extracts structured attributes before any LLM call, so generation works from clean facts rather than raw spreadsheets. Fine-tuning plus tone modulation handles channel adaptation; conversion analytics close the feedback loop.

Personal finance tools either bucket transactions with brittle rules or hide everything behind opaque scores — neither predicts where spending is heading, and few treat financial data with real security discipline.

Hybrid inference: deterministic rule-based categorization for unambiguous transactions, LSTM recurrent forecasting for expense trajectories and financial-health scoring, statistical fusion models for budget recommendations.

Time-series forecasting on sparse, irregular personal transaction data; separating genuine anomalies from one-off purchases; storing highly sensitive financial data safely.

The rule layer supplies clean labeled sequences to the LSTM, improving forecast stability on small datasets. Anomaly alerts compare predictions against observed spending, not fixed thresholds. Security is layered: AES-256 storage, compartmentalized process execution, hashed identity mapping.

Sales teams leak leads because CRMs record activity but don't reason about it — pipeline prioritization, churn detection, and follow-up routing all stay manual.

Next.js frontend over microservice-aligned backend services with a plugin-ready AI inference layer: Llama 3.1 for multimodal lead intelligence, predictive scoring for pipeline ranking, unsupervised clustering for segmentation, behavioral profiling for churn.

Combining LLM-based lead understanding with classical ML scoring in one coherent pipeline, and keeping the inference layer swappable so models upgrade without rewriting services.

Each intelligence concern — scoring, segmentation, churn, routing — is its own service behind modular API endpoints, with the AI inference layer as an explicit plugin boundary. Automated communication routing acts on model outputs without manual triage.