Samuel Kahn

Leadership

I built and run the roughly twenty-person organization behind RiskWise: applied AI scientists, ML engineers, data scientists, software engineers, and product. I started as Tickr's first AI hire, grew the function from a team of one, hired most of it, and set its structure. As AI became the center of the product, engineering and product were folded under me alongside it, and I now own the technical direction across all three.

The job is mostly judgment and people. Ten or more initiatives run at once, and I decide which ones get the compute and the headcount. I set research strategy, hire, develop and promote from within, and keep the team close to customers so they understand what a failure actually costs. I carried the group through a full pivot to AI-first products with the team intact, and kept it shipping the whole way.

Outside the org chart I'm the technical face to enterprise buyers, from the Fortune 10 to the Fortune 500, and I negotiate the partnerships the platform runs on: data providers, labeling vendors, and model platforms including Anthropic and OpenAI. I've also acted as fractional AI lead for three startups, two of which were acquired.

I stay close enough to the technical work to lead it credibly, still hiring, reviewing architecture, and getting into the code. Most of my impact, though, is in setting direction and then handing over real ownership. The thirty-task deep-research evaluation, for instance, was a direction I set and a scientist on the team carried end to end, which is how I want it to work.

Agents, post-training, and evaluation

Most of my hands-on work is on agent systems and post-training. I post-trained a 4B risk-attribution model that cut errors roughly 50% against GPT-5.2 at nearly 20× lower inference cost, and fine-tuned Qwen 3 with QLoRA/PEFT on eight H100s alongside bge-large-en-v1.5 for retrieval, reaching 98.1% F1 across four hierarchical datasets against 86.1% for GPT-4o and 86.6% for Claude 3.5 under identical conditions. I've built agentic discovery using a Thompson-sampling bandit over natural-language research briefs run as frontier-model search episodes, with reward computed in code rather than by the model, and an MCTS-guided research system that attacks its own findings before recording them. On evaluation I've stood up adversarial LLM-as-judge, capability, regression, and human-gold suites for actuarial, legal, and compliance review. New production failures start as capability tests and, once reliable, become regression tests, so a later model, instruction, or system change cannot quietly undo them. I set the strategy for a 30-task evaluation of high-stakes risk deep research and mentored the scientist who ran it; against open-web retrieval it improved source traceability by 10.4 points and analytical rigor by 3.3. agent infrastructure · RAG tuning · evaluation

Risk intelligence

I lead RiskWise, an agentic risk-intelligence platform licensed by major insurers and Big Four firms. It continuously monitors a catalog of tracked companies, risk topics, and industries, ingesting from more than 50 licensed and public sources: LexisNexis, CourtListener and RECAP, JPML dockets, SEC EDGAR, DOJ enforcement actions, FDA recalls and FAERS, PubMed, news, X, and Reddit. The work is framed around legal and property-and-casualty exposure, including product liability, D&O, transport, antitrust, cyber, and employment practices, and looks for the correlation risk, tail risk, and emerging threats that conventional taxonomies miss.

Turning that volume into signal is most of the engineering. A processing pipeline disambiguates entities, filters spam, reranks passages for relevance with a fine-tuned embedding model, and decomposes each risk into drivers whose movement is tracked as an index over time. On top of it run analyst agents grounded in that evidence: a deep-research agent that writes underwriting-grade reports with citations back to source, company- and driver-level narrative summaries, and a risk search agent. Each one cites the evidence it drew on so an underwriter can trace a conclusion back to the filing or article behind it. The output feeds underwriting, portfolio diversification, and board risk.

One of the predictive layers forecasts litigation. A multi-armed bandit over research agents feeds a discrete-time hazard model of multidistrict-litigation consolidation at 0.90 held-out AUC; backtested point-in-time on data frozen at January 2022, 2023, and 2024, between 16 and 32% of its top 25 ranked risks later became MDLs against a base rate under 1%, surfacing them 8 to 24 months early. benchmarks against frontier models

Legal AI

Seven of my thirteen years in AI have been in legal. At Subpoena Solutions I built the proof-of-concept of an intake system that reads subpoenas, search warrants, Section 2703(d) orders, preservation demands, and emergency requests, classifies the demand, and drafts grounded objections. Six LLM classifiers are grounded in 50-state law. I used DSPy and MIPROv2 to improve them, then added an embedding-based normalization layer so an incorrect generation could not break the downstream workflow. I've fine-tuned models on attorney-reviewed legal process data and built evaluation with confidence-gated attorney review, cutting errors more than 50% against frontier baselines on 20+ tasks. For privacy-sensitive clients whose policies prohibit fine-tuning, I've mentored engineers building self-improving agentic loops that optimize instructions rather than weights. Two of the three legal-tech startups I've advised have been acquired.

The role has been as much leadership as engineering. At Subpoena Solutions my work led to a private-equity acquisition and to standing up a dedicated AI and engineering team; I run weekly mentorship for that team and advise its leadership on staying differentiated as general models absorb more of the work. At SpeedLegal I worked with the CEO on that same question and defended the technology to investors during a contested fundraise. At Joinder, since acquired by Brightflag, I built the signature-detection and legal-taxonomy models the product shipped on. Across RiskWise I'm the lead architect for the legal-risk features and set the direction the team builds against.

Predictive modeling and forecasting

A decade of demand forecasting and causal inference on CPG and retail data. I've built SARIMAX and Prophet pipelines with exogenous regressors, holiday calendars, and Bayesian MCMC fits, benchmarked univariate against multivariate specifications with automated hyperparameter search, and run these at retailer and product-group granularity for national brands. I built Generative Predictor Search over millions of economic series using retrieval, reciprocal rank fusion, and LLM reranking, which cut out-of-sample MAPE 15.6% and ran 13× faster than exhaustive search. I hold a US patent application on causal inference for marketing campaign impact, and trained DeBERTa-XXL across 500+ product classes with a triplet-loss embedding model that anchored a multi-million-dollar contract. My work has crossed financial services, insurance, healthcare, legal, and science: for FIS I worked on dynamic product categorization and applied early reasoning LLMs to Medicare Advantage benefit classification. GPS · causal impact

AI for science

I hold a research fellowship at UC Santa Cruz alongside the CTO role. I'm first author on NEO, a conditional GAN that improves the accuracy of physical measurements by factors of 2 to 10, with code and weights released publicly and coverage in the NVIDIA blog. I'm a co-author on a Nature Portfolio paper, and I built an autonomous discovery system that ran 49 analyses across 25 years of satellite data and reported one of its three hypotheses as inconclusive rather than forcing a conclusion.