Bevra: agentic content engine behind a webhook seam
Multi-tenant SaaS calling a self-hosted agentic backend across three webhook endpoints.
A deterministic eval gate and an offline LangSmith harness read the same criteria.py.
Applied ML, experimentation, and product discovery.
Ten years turning mixed-source data into quantified opportunity spaces and business cases. Behavioural, telemetry, unstructured, market. The case studies below carry the working artefacts: the eval-gate criteria file, the JTBD graph, the metrics framework.
Applied ML
Causal Inference
Experimentation
LLM + RAG
JTBD Discovery
Business Cases
Featured work
Kosovo #MoseMerrMalin: JTBD study for a reforestation campaign
Five B2C segments. Core Job, Big Job, and TAM/SAM/SOM sizing per segment. Landing-page copy shipped into the client's website.
See the project →Giro555: metrics framework for a national charity alliance
Primary, supporting, and counter-metrics per test. Decision dashboard readable by a non-analyst stakeholder.
See the project →Methods
Opportunity discovery
Jobs To Be Done, segmentation, TAM/SAM/SOM sizing, risky-assumption analysis, discovery workshops.
Experimentation
A/B and multivariate testing, primary / supporting / counter-metrics, cohort and retention analytics, qual and quant integration.
Applied ML and causal
XGBoost, time-series forecasting, scenario and portfolio forecasting, stochastic simulation, uncertainty quantification, DiD, synthetic controls, IV.
LLM systems
Retrieval-augmented generation, structured output, deterministic eval gates, golden-set evals, guardrails, human-in-the-loop.