AI security · · 2 min read
Designing Teger AI to explain phishing
How Teger AI, an explainable social-engineering defence for Gmail and Slack, pairs deterministic signals with AI analysis and treats every message as untrusted.
Emmanuel Effiom DukeAI engineer
LLM products, agents and evaluation tooling, built end to end and shipped.
I build AI systems that have to work outside a notebook: products with real users, adversarial inputs, API budgets and failure modes that don't announce themselves. My approach is the same whether it's a security product or a forecasting pipeline. Separate what can be checked deterministically from what needs a model. Treat model output as something to validate, not trust. Measure before tuning.
What I build
- LLM products. Teger AI pairs deterministic signal extraction (urgency, credential and payment requests, authority claims, URL shape) with AI forensic analysis. It validates every result against a typed schema and treats each message as attacker-controlled input.
- Agents and multi-model pipelines. My Metaculus forecasting bot runs parallel retrieval across several search providers, synthesises a research brief, and combines a weighted five-model ensemble with critic and red-team passes.
- Evaluation and reliability tooling. An offline harness that replays calibration variants against 190 resolved questions at zero token cost. A leak canary for date-bounded backtests. A probe that measures each model's real knowledge cutoff instead of trusting its self-report.
- Full-stack AI applications in fintech, agriculture and geospatial analysis, from backend APIs to the interface.
How I work
Most AI failures I've debugged were silent: a research provider failing without crashing, a quality signal stuck at its maximum, a calibration step biasing every forecast the same way. So I log what each component actually did, keep secrets server-side, rate-limit public endpoints, and build small, cheap evaluations before scaling anything. I write up the results, including the negative ones.
Tools
Python, TypeScript and JavaScript, FastAPI, React, Chrome extensions, OpenRouter and multiple model providers, retrieval and search APIs, Pydantic schemas, GitHub Actions CI, and PyTorch for research work.
Selected work
- Teger AIExplainable social-engineering and phishing detection for Gmail and Slack: Chrome extension, FastAPI backend and React security console (MVP v0.2, private beta).
- oracledeckv1 forecasting botMulti-source research, a five-model LLM ensemble, critic and red-team passes, and an offline evaluation harness for Metaculus AI tournaments.
- LLM knowledge-cutoff probeMeasures nine models' real training cutoffs by dated-event bisection.
- AxiosPayCross-border FX for African travellers; 4th place, Interswitch × Enyata Developer Hackathon.
- FarmVoice-AIAgricultural extension AI agent for farmers.
- CropieSatellite-powered crop-disease early warning.
- GeoDrillAI drilling-probability scoring with a live risk heatmap.
Related writing
Forecasting · · 3 min read
Building a Metaculus forecasting bot
Lessons from an LLM forecasting bot in Metaculus AI tournaments: silent research failures, extremized toss-ups, and replacing vibes-tuning with measurement.
AI evaluation · · 2 min read
Ask the calendar, not the model
Measuring nine language models' real knowledge cutoffs with dated-event bisection: every model that self-reported a cutoff was wrong, some by over a year.
Other areas
Expertise
AI safety researcher
Mechanistic interpretability, model auditing and alignment experiments, with open code and honest negative results.
Expertise
AI security engineer
Social-engineering and phishing detection, prompt-injection-aware LLM design, and auditing models for hidden behaviour.
Expertise
AI forecasting and LLM evaluation
Automated forecasting with LLM ensembles, calibration measured offline, and probes that test what models really know.
Expertise
Mechanical engineer and energy researcher
CFD modelling, machine-learning energy optimisation and air-quality research, grounded in a mechanical-engineering degree and hands-on HVAC work.