Reasoning-First Architecture

In 2026, HR-tech is making almost the same bet, almost unanimously: LLM agents will replace recruiting matching by 2027. Josh Bersin frames it as the top layer (Layer 4) of the talent-intelligence stack. The bet is obvious and near-consensus.

I'm betting against it.

Not because I'm against AI — I work with AI every working day, and in the product I'll describe, every line of code was written through Claude, not one by my own hand. But because 135,000+ vacancies and 1.1M Telegram messages that have passed through my system since April 2026 have shown me the opposite: a deterministic reasoning-first core + LLM as augmentation (not replacement) wins on precision, latency, cost, and EU AI Act compliance. This essay is about why.

Who I am and why I'm writing

I started in SEO in 2006 — twenty years of watching how content + structure + retrieval algorithms shape outcomes for billions of queries a day. HR-tech matching is built the same way: content (descriptions, CVs) + structure (roles, stack, levels) + retrieval → outcome.

TD-CRM v2 is a recruitment CRM for the CIS-IT market: 135,000+ vacancies (61,000+ open), 1.1M processed Telegram messages, ~1,990 monitored channels, 14 directions, 15 countries. Publicly self-classified as high-risk under the EU AI Act, Annex III #4. And not built by an engineer — I don't write code. I set architecture and product; Claude generates the implementation — the pattern Karpathy called "agentic engineering" in March 2026. Replit reached a $3B valuation in September 2025 on the same pattern. I'm not the first to build this way; I just state openly that I do.

Here's why I decided to write against the near-consensus. When you look at 135,000 vacancies not from above but hands-on, something plain emerges: the overwhelming majority of matching is decided deterministically — stack, grade, format, geo, salary band. LLM is genuinely needed for parsing messy text, not for the decision itself. Yet the industry puts the LLM into the decision. Twenty years in SEO taught me how that ends when you place a fuzzy layer where you need reproducibility and explainability. For me this isn't a theoretical argument about trends — it's what I see in my own data every day.

Where the industry's bet is wrong

LLM-first matching is the wrong bet on four axes: precision, cost, regulatory compliance, traceability.

Precision. LLM-fuzzy matching carries built-in variance call to call — the same candidate/vacancy pair scores differently between runs. Fine for a chatbot; a defect for matching where the decision becomes part of an audit trail.

Cost. An LLM call per attempt × thousands of matches a day = thousands of dollars a month a pre-revenue company can't carry. Deterministic field-matching is near-zero marginal cost.

Regulatory compliance. EU AI Act, Annex III #4 enters enforcement in August 2026. High-risk systems must log decisions with the basis for each. LLM-first log: "matched because the model said so." Reasoning-first log: "stack overlap 8/12, Senior (7 years), Remote (true), salary overlap. Score 0.84." One of these exists in an audit trail. The other doesn't.

Traceability for DD. Acquirer's counsel asks "how exactly was this matched?" LLM-first is opaque; reasoning-first is explainable by construction. Market signal — Eqtble → Paradox (Feb 2025): people-analytics without trading-grade data classifies as HR-tech. Same logic: LLM-fuzzy without traceable rules isn't serious HR-tech.

A caveat against a straw man: LLM isn't "bad." LLM in the wrong layer is what's bad.

What the right architecture is

Reasoning-first means any decision that must be defensible — matching, scoring, ranking — goes through deterministic rules first. LLM is used ONLY where no deterministic rule is possible (free-form parsing), where uncertainty is explicit (confidence score + degraded path), and where output doesn't route directly into a final decision. In TD-CRM the matching score is an explicit formula (stack overlap + grade + role + work format + salary), not "model, rate the fit."

Monolith over micro-services, deliberately: one founder + Claude = no team to carry distributed-systems overhead; a deterministic core needs a single source of truth. Per ADR-0001 — a single FastAPI + PostgreSQL + Redis + arq worker + SSR. The "Modern Monolith" in the spirit of DHH / 37signals — coherent with reasoning-first, not contrarian for its own sake.

Per decision D10 the system publicly self-classifies as high-risk (Annex III #4): log of decisions (Loki + structured logging), risk management, human oversight (a recruiter reviews every match, no auto-action), data governance. Self-classification pre-empts a DD finding; honest disclosure builds trust.

Coverage-gap: a vertical pipeline as an asset

LinkedIn has been blocked in Russia since November 2016. Lightcast, Revelio Labs, People Data Labs source most data from LinkedIn; their coverage east of the Riga line drops by roughly 40% (my estimate, from their public statements). Structural, not temporary.

Instead — a Telegram-MTProto pipeline: ~1,990 monitored CIS-IT channels, 1.1M processed messages. Data global vendors physically can't get (sanctions, blocked services, technical access). Marc Casado (a16z, 2019, "The Empty Promise of Data Moats") showed data scale alone isn't a strong moat; pipeline-IP is harder to replicate. Our CIS-IT focus is a differentiating asset, not a forced weakness.

Agentic engineering: who built this, and how

Directly: TD-CRM v2 is built 100% through Claude. I don't write code — I direct architecture and product; Claude generates the implementation. Karpathy split the pattern into "vibe coding" (raises the floor for non-coders) and "agentic engineering" (preserves the quality bar). The difference is rigor evidence: 14 ADRs in docs/adr/, observability (Loki + Grafana + Prometheus), weekly dependency scanning (ADR-0013), a public changelog from Conventional Commits, contractor review of security/PII paths before a deal.

Replit reached $3B in September 2025 on this pattern (ARR $2.7M → $70M in 12 months). Market-validated. An acquirer reads this as "a founder with rigor + AI leverage," not "vibe-coded something" — the difference is artifact density: this ADR series, the changelog, the runbook, observability, and this essay.

The acquirer angle

If you're a corp-dev analyst at Greenhouse, Workable, iCIMS, or Lightcast — this is for you too. Directly: TD-CRM is pre-revenue, 0 paying customers, $0 ARR. Not a "team + customer book" target — three separate assets.

First — pipeline-IP: 135,000+ vacancies, 1.1M messages, ~1,990 channels since April 2026; data Lightcast/Revelio physically can't collect. Building an equivalent = 18-24 months + sanctions risk + Telegram-API expertise. Second — reasoning-first matching/scoring code, reproducible in the acquirer's infra. Third — the Estonia OÜ entity (Q2 2026), EU jurisdiction, no origin complications.

Why reasoning-first matters to the buyer: EU AI Act enforcement from August 2026 hits LLM-first products with compliance-retrofit pain. Reasoning-first is already compliant and shipping — acceleration, not a retrofit. What the buyer doesn't get: customer base (0), brand, team, ARR (0) — not priced in. The sum of the three assets is below $15M.

Closing

This is a bet: one solo founder + Claude as agentic engineer + a few weeks = a reasoning-first architecture + a CIS-IT pipeline + 14 ADRs. The bet is that this combination outperforms LLM-first global aggregators in HR-tech matching by the end of 2027. I'm publishing now — to fix the bet before the result, not after.

If you have a reasoned critique, write. Honest pushback is worth more than silent doubt. And no "request a demo" at the end: if the essay isn't convincing on its own, a demo won't fix it.