A capability knowledge base and RFQ parser let an Agent classify each requirement as compliant, non-compliant, or uncertain, cite evidence, and fill the original Excel template.
SDMC (Shenzhen SDMC Technology), founded in 2003, is a national high-tech and "little giant" enterprise combining hardware, software and cloud services for the smart home, organised around home connectivity, home entertainment and whole-home intelligence.
Anyone who has worked on a tender knows the punishing part is not writing the solution narrative — it is checking the customer's technical requirements line by line. An operator's RFQ arrives as an Excel workbook running to several hundred rows, each one asking: can your product do this? which standards does it support? do you hold the certification? can you meet the delivery terms?
And the answer cannot just be compliant or non-compliant. Every judgement needs evidence behind it — a page of a product specification, a table in a chip datasheet, a paragraph from a past project, a clause in a certification document. Without evidence you cannot conclude; get one line wrong and it costs scoring points; miss one line and you have buried a risk in the bid.
In set-top boxes, Android TV and smart terminals the requirements are hard-edged: SoC, video codecs, DVB / IP, secure boot, 4K HDR, Dolby, Launcher, broadband gateway, operator customization. A model's general knowledge cannot answer these. Phrases like usually supported or generally possible are high-risk language in a compliance setting.
So a single tender occupying several people for three to five days is unremarkable. Worse, the judgement that matters sits with a handful of senior staff: newcomers cannot take it on, and when the experts are busy the whole process becomes a bottleneck.
The two teams handed that stretch of the process to a bidding agent. Not a helper that writes answers, but a re-organization of the most repetitive, most time-consuming and most error-prone part of tender response into a workflow that can be verified, traced and improved.
Five steps: upload the customer's Excel template; the system identifies which columns are item number, requirement and response, and extracts the real requirement list out of several hundred rows; the agent works through it against the company knowledge base, judging compliance and attaching cited evidence; where information is insufficient it asks a person; after human review the answers are written back into the customer's own Excel, ready to deliver.
The critical engineering constraint is tool ordering: the agent must submit evidence before it is allowed to judge. It cannot answer off the cuff. Each judgement states what it searched for, which passage it found, and which document and page it came from.
When the knowledge base holds no direct evidence, the agent does not fill something in — it stops and asks. In a tender, uncertain but carrying on is far more dangerous than uncertain and asking. The system also keeps a multi-modal fallback: when the parsed text is unclear, the agent opens the original PDF image and table directly, so vision recovers what text extraction dropped.
An accompanying operations and observability layer records every model call, retrieval hit, tool use and per-item latency, so any change can be regression-tested against a fixed gold-standard set.
On an internal gold-standard benchmark built from the customer's real tenders — a 78-question template with matching specifications and chip datasheets in the knowledge base — the agent reached 92.5% answer accuracy.
What matters more is that the number was reached by iterating on the same benchmark: 75.7% → 84.9% → 92.5%. The team defined a 4+2 measurement set (question identification, answer accuracy, coverage, per-item time, plus post-clarification accuracy and coverage), so whether it improved, and by how much, is a number rather than an impression.
A tender that used to take several people days of cross-checking now runs end to end in tens of minutes.
The human role changed too: from finding, judging and filling line by line, to reviewing the agent's answers, supplying the judgement calls and controlling risk.
And the experience that sat with senior staff is becoming reusable knowledge base content, prompts, tool flows, gold samples and review rules — which is the real value of an agent entering an enterprise process. Not answering a few questions once, but turning experience into system capability.