Case file 01 / Independent Project
OptiTalent
Evidence-first recruiting workflows
A multi-tenant recruitment product that turns candidate search, evidence review, shortlist decisions, controlled sharing, and recruiter feedback into one coherent workflow.
- Classification
- Independent Project
- Delivery status
- Independent product / active validation
- Case
- 01 / 07
The problem
Recruiting teams need to move quickly without turning matching into an opaque score or losing the evidence behind a decision.
System arrangement
How the system works
Confirmed role constraints fan into exact and semantic recall lanes. Candidate-level RRF builds the pool, ColBERT refines it, and evidence policy controls what can reach a recruiter shortlist.
- 01SourceRole intent + constraints
Confirmed eligibility compiles first
- 02CapabilityPGlite exact + lexical
Canonical facts, filters, and FTS
- 03CapabilityQdrant dense + sparse
BGE-M3 versioned derived index
- 04CapabilityCandidate-level RRF
Fuses only executed recall lanes
- 05CapabilityColBERT MaxSim
Reranks the fused candidate pool
- 06ControlEvidence policy
Unknown stays unknown, not excluded
- 07OutcomeRecruiter shortlist
Evidence remains reviewable
- 08ControlTenant-local compute
Redacted subjects stay close to data
My role
I defined the product boundaries, tenant model, evidence contracts, recruiter workflow, review gates, and release acceptance. AI agents produced most implementation code under those constraints.
Capabilities
- 01
Hybrid lexical and dense candidate retrieval
- 02
Evidence-preserving reciprocal-rank fusion
- 03
Local embedding and Qdrant recall runtime
- 04
Candidate evidence and explainable match signals
- 05
Draft and approved shortlist states
- 06
Controlled sharing and feedback
- 07
CLI, MCP, desktop, and web access paths
Product decisions
- 01
Prefer evidence-first retrieval and explicit unknowns over a single opaque LLM judgment.
- 02
Place bounded embedding and retrieval work close to tenant data while preserving server-side policy and review authority.
- 03
Keep the policy and lifecycle authority behind stable contracts while interfaces render approved state.
- 04
Treat agent and human entry points as adapters over the same capabilities.
Evidence
- •
Public product surface at app.optitalent.cc.
- •
Implemented candidate search, candidate evidence, shortlist, and sharing contracts.
- •
Tenant-scoped retrieval and explicit review states are present in the product runtime.
- •
LightGBM LambdaRank was evaluated against deterministic baselines on a frozen retrieval protocol and was not selected as the production winner for that feature shape.
The design principle
Recruiting software often compresses a complex judgment into a rank. OptiTalent keeps the underlying evidence visible: what matched, what remains unknown, which records support the conclusion, and where a recruiter still needs to decide.
Capability shape
The product is organized around a recruiting project rather than a loose collection of AI features. A team can describe a role, inspect the structured project, search its private talent graph, build a draft shortlist, approve candidates, share a controlled view, and collect feedback.
The same capability boundaries support human and agent entry points. Search, candidate evidence, shortlist construction, and sharing are callable contracts, not logic hidden inside one page.
Retrieval maturity
Recall is treated as a product system rather than a vector-search checkbox. PGlite exact fields and lexical search provide deterministic anchors. Qdrant is the retrieval engine for the rebuildable tenant-scoped projection, combining dense, sparse, and ColBERT-style late-interaction representations. Candidate-level reciprocal-rank fusion preserves evidence across executed lanes so a reviewer can inspect why a candidate entered the set.
Bounded embedding and recall work can run close to tenant data, reducing unnecessary transfer while server-side contracts retain policy, lifecycle, and review authority. Learned ranking is measured rather than assumed. LightGBM LambdaRank was evaluated on a frozen protocol, but it is not presented as a production winner where the benchmark did not justify the added complexity.
What I learned
The useful unit of AI product design is not a prompt. It is a capability with inputs, state, permissions, evidence, failure behavior, and a place for human judgment.