The build · Applied AI

Matching, scoring and retrieval — reaching production, staying there.

Deliberately narrow. We don’t sell AI strategy and we don’t build demos. Every system here produces a defensible score, a ranked shortlist or a grounded answer, and gets used in an operational decision the day it ships — which means it has to survive contact with real data, real edge cases, and someone asking why it said what it said.

01

How we think about it

Four principles, before any architecture diagram.

01

Grounding over generation

A model doesn’t get to assert something the underlying data can’t support. Every output traces back to a retrieved, validated source or a computed feature.

02

Explainability by default

If a score or a match can’t be decomposed into the factors that produced it, it doesn’t ship. “The model said so” is not an answer a client’s own team should ever have to give.

03

Evaluation before deployment

A system gets measured against held-out, labelled outcomes before it touches a real decision — not monitored into correctness after launch.

04

Human judgment at the edges

High-confidence cases get automated. Low-confidence cases get routed to a person, with the system’s reasoning attached.

02

Decisions & scoring

Turning a pile of candidates into a defensible decision.

The two capabilities clients call us for first — a ranked shortlist, or a risk flag before it costs them something. Both run on the same discipline: weighted, calibrated, explainable.

Candidate & entity matching

Scoring, not filtering

Keyword search returns anyone who used the right words. We build weighted multi-factor scoring — relevance, potential and likelihood-of-fit scored separately, then combined — calibrated against real historical outcomes, with every ranked result explainable down to the factors that produced it.

Proven — EY India, Scalekrut
30+scoring parameters, calibrated against outcomes
Risk & attrition scoring

Forward-looking, not a lagging report

A dashboard that tells you who already left is a historical record. We build on leading signals — tenure patterns, engagement trends, manager-change events — with thresholds tuned against the real cost of a false positive versus a missed case.

Same scoring discipline as matching
2failure modes engineered against: false positives, stale signals

03

Knowledge & retrieval

Representing what’s related, and answering only from what’s true.

A graph for the questions a relational join can’t answer cheaply. Retrieval for the questions a model shouldn’t answer from memory alone.

Graph-based relationship modelling

Why a graph, specifically

A relational schema answers “does this row match that row.” It can’t cheaply answer “who is one hop away from a strong match” — usually the question that actually matters. We reach for a graph only where the query shape needs it; transactional records stay relational.

Proven — Scalekrut
1hop-of-adjacency, found by walking the graph — not a join
Retrieval-grounded generation

The full pipeline, not just “we use RAG”

Content chunked for how it’s actually queried, embedded into a tenant-isolated vector store, retrieved under a strict tenant boundary, then assembled into a grounded prompt. When nothing relevant is retrieved, the system says so instead of generating a fluent guess.

Built on Qdrant, tenant-isolated
D1Ingestion &chunkingContent broken intoretrieval-sized unitsD2Embedding &indexingVectors in Qdrant,tenant-isolatedD3RetrievalConstrained bytenant & contextD4ContextassemblyGrounded, not themodel's own priorD5Confidence &refusalNo support found —the system says soNo retrieval match at D3 — path exits at D5 with a stated refusal, not a guess.

The retrieval-grounded generation pipeline — five stages, refusal built in rather than bolted on.

04

Autonomy

Beyond a single call to a single model.

Where a scoring engine or a RAG pipeline isn’t enough — a task genuinely needs multiple steps, tool use, or a model that speaks a domain’s vocabulary natively. Built and specified in real depth; not yet proven on a delivered engagement, and flagged as such rather than blurred.

Agentic orchestration

Beyond a single call to a single model

A goal decomposed into steps, each with a scoped tool and a defined success condition. An agent gets exactly the tools its task requires — irreversible or high-stakes actions stop at a human approval gate before they execute, with every step logged and reviewable.

Capability built · not yet delivered
Fine-tuning & domain-adapted models

Only when prompting is the wrong tool

Parameter-efficient fine-tuning (LoRA/PEFT), evaluated against a prompted baseline before it ships — or, further still, continued pretraining of an open-weight model against a client’s own domain corpus, for when the model needs to have absorbed the domain, not just been steered toward it.

Capability built · not yet delivered
01TaskdecompositionA goal broken into steps,each with a scoped tool02Scoped toolcallExactly the tools thisstep needs, nothing more03Humanapproval gateIrreversible or high-stakesactions stop here first04Audit trailEvery step and result,logged and reviewableFine-tuning and domain-adapted models sit alongside this loop — used when repeated prompting is the wrong tool.

The orchestration loop — a human approval gate sits between the agent and anything irreversible.

05

Evaluation

A model that hasn’t been evaluated against held-out data isn’t a finished system — it’s an untested hypothesis with a UI in front of it.

Offline evaluation

Scoring and matching logic tested against labelled historical outcomes before it ever sees a live decision.

Precision & calibration tracking

Not just “is it accurate” — is a confidence score of 80% actually right eight times out of ten, tracked over time.

Human-review sampling

A sampled slice of automated decisions gets checked by a person on an ongoing basis, not just at launch.

Drift monitoring

The data a model sees in month six is rarely identical to what it saw at launch. We watch for that shift rather than assuming day-one performance holds.

06

Quality engineering

A separate capability, also AI — applied to engineering itself.

Different from everything above. Candidate matching and RAG are AI applied to a business decision; this is AI applied to the tests that guard a system — generating and maintaining coverage, spotting the edge cases a manual test plan misses. Applies to any stack, not only the legacy environments it’s often paired with.

AI-assisted test generation

Coverage proposed, not hand-written

Coverage proposed from the actual code paths and data shapes in a system, rather than written by hand against what a tester assumes it does.

Capability built · not yet delivered
Continuous test maintenance

Suites that don’t rot

Suites flagged and updated as the underlying system changes, instead of quietly rotting until a release breaks something nobody was testing for.

Capability built · not yet delivered

07

The stack

Mainstream infrastructure underneath all of it.

Java  ·  Spring Boot  ·  Node.js  ·  Python  ·  React  ·  Neo4j  ·  Qdrant  ·  Microsoft Azure, client-hosted

Next

See the platform underneath — Strata.

StrataBook a 15-min architecture call