Built for a multi-business-line enterprise with a heavy tech-hiring funnel, the system combines LLMs, multi-stage filtering and semantic matching to handleresume parsing, candidate scoring and role matching — with decision support for every step.
In late 2025, this multi-business-line enterprise client kicked off another tech-hiring push. The first instinct was the obvious one — hire more recruiters. It didn't help.
They were burning hours just organizing them. The real screening problems only surfaced after that.
The JD says “high concurrency, microservices, distributed cache.” The candidate writes “flash-sale system, Spring Cloud refactor, Redis hot keys.” To engineering they're the same skill. To a non-technical recruiter, there's no way to make that call in 30 seconds.
Same role: recruiter A weighs degrees, B weighs FAANG-style backgrounds, C counts keyword hits. The interview pool ends up wildly inconsistent.
First time the client realized: their hiring bar was never actually structured.
Search “Go developer” and resumes saying “Golang / cloud-native backend / microservices” never surface. Search “user growth” and “Growth / growth ops / viral strategy” gets dropped too.
ATS finds “resumes that look like the JD,” not “people who actually fit.”
Teams kept saying “great people never even reached us.” Engineering managers ended up doing first-round screening themselves — recruiter throughput dropped, engineers duplicated work, time-to-hire kept stretching.
At worst, time-to-first-review on a Java role hit > 72 hours.
An LLM-only approach falls apart in production for three predictable reasons. That call shaped everything that came after.
Same resume, different time, different prompt — scores drift. Recruiters can't build trust, and trust is the first gate any AI tool has to pass.
Recruiters don't ask “why is this an 82.” They ask “why is this person ranked above the next one.” Without traceable skills, experience and role-fit, they'll re-screen by hand anyway.
With a six-figure resume pool, full LLM inference on every candidate means latency, token spend and concurrency all blow up. The AI becomes the new bottleneck.
Let each layer do what it's good at: rules for stability, vector recall for coverage, LLM for context. AI no longer has the deciding vote — it's one input into the scoring system.
Engineering isn't just calling an LLM. We broke the recruiting flow apart so each module owns what it's best at — and AI stays controllable, explainable and trustworthy.
The easiest part to overlook. The first step isn't AI — it's RPA + OCR + schema normalization: pull from every platform, dedupe, parse PDF / Word / images into one canonical structure.
When the JD says “high-concurrency experience,” the system understands the connections between Redis, Kafka, flash-sale systems, distributed cache and service governance. It matches on “has this person actually done this kind of work,” not “does the word appear.”
We started with the model emitting raw scores. They drifted, they were unstable, and they didn't explain themselves. We moved to rules-led, AI-assisted: weighted scores across skills, experience, contextual depth, logical structure and role fit. AI handles context; rules make the call.
When a candidate says “I optimized a high-concurrency system,” a non-technical recruiter has nothing to follow up with. The system generates the follow-ups automatically: “how would you handle cache penetration?” “why did you partition Kafka that way?” — questions that actually probe.
What the client said afterwards:“Engineering managers used to help HR run the first screen. We can finally split that work apart.”
We ran a real (anonymized) resume through the full pipeline: 3 pages of PDF, messy formatting, conversational phrasing — under 10 seconds end to end.
Pulls out skills, projects, timeline, education and work history.
Maps “Redis + Kafka optimization” to distributed cache, message queue, high-concurrency governance, then vector-matches against the role.
Weighted across skills, experience, contextual depth, logical structure and role fit — rendered as a radar.
Technical probes + project probes + architecture questions + risk questions — handed straight to the interviewer.
Java backend · 5 years · distributed experience match
From ingestion and talent pool, to semantic search, deep analysis and interview-question generation — five core screens, mapped to the five pipelines in the case.
Outlook, Gmail, NetEase Mail, QQ Mail and WeCom — 5 source types connected in one click. RPA keeps watching for new mail, parses attachments, normalizes fields and feeds them downstream. This is where the “we can't even collect the resumes” problem actually gets solved.
These problems are harder than “call an LLM.” They decide whether an AI system actually survives in production.
For the first time, the whole recruiting flow worked as a real human + AI loop.
We handed “read the first round of the world” to the system — rules for stability, models for understanding, scores that explain themselves.
Recruiters went back to the work that matters — talking to people, judging fit, moving the conversation forward, spotting risk. The part AI doesn't replace.
“Recruiting isn't keyword filtering.
It's about understanding people.”
What actually worked wasn't “AI replaces HR.” It was “AI reads the first round of the world for HR.”
Related cases
Other projects delivered by Wavesteam — AI, IoT, platform builds and enterprise software.
A donation and student-aid platform for non-profit organizations — platform architecture and user flows tailored to charity operations.
View case
A continuous livestream and audience-engagement platform built for environmental campaigns — covers production flow, scenario orchestration and post-event capture.
View case
End-to-end tournament registration, real-time scoring and an achievement system that brings golf tournaments online.
View caseThese projects usually involve business systems, device integration, AI workflows or multi-role back-offices. We assess feasibility against real delivery constraints and give recommendations close to the implementation stage.
